Think you can build real AI? Prove it. A student-only program to discover and hire our next generation of AI Builder Interns.
Students only. 6 or 12 month AI Builder Internship. In-person, Bangalore, from September.
No resume screening. No long application. Four steps: pick a track, build something real, show your work (a public repo, a 5 minute pitch video, the architecture), and if it has signal we call you in.
Grow the merchant’s revenue, and make them sellable to AI buyers.
Build an agent that grows revenue for a merchant on Razorpay test-mode APIs, or that makes a merchant transactable by an AI buyer end to end.
Why now: NPCI’s UAP and the global protocol race (ACP, AP2, x402) make agent-to-agent commerce the open problem of the year, and Razorpay’s in-app pilots are already live.
Example directions: Conversational in-app checkout, Agent-readable catalog, Upsell & cross-sell agent, Campaign orchestrator.
The bar: Every money action explainable, bounded and gated. Show the audit trail and one failure handled gracefully.
Stop the merchant losing money to fraud, returns and chargebacks.
Build a working detector, verifier or auto-responder for one class of loss, with measured precision and recall on a held-out test set.
Why now: AI-enabled fraud is hitting Indian BFSI while returns and chargebacks quietly eat margin. This track surfaces the risk and ML minded builders the others miss.
Example directions: Chargeback evidence responder, Return-risk scorer, Fraud-spike detector, Abuse-ring sentinel.
The bar: Honest metrics including false-positive cost. Strictly defense-only: anything offense-capable is disqualified.
Find revenue that’s slipping away and win it back.
Build an agent that detects revenue at risk, determines the right intervention, and executes a bounded recovery workflow: from payment failures and checkout abandonment to overdue receivables.
Why now: Revenue loss rarely happens in one clean step. A payment degrades, a checkout gets abandoned, a subscription fails, or an invoice goes overdue. AI can now close the loop from detecting the problem to diagnosing it, choosing the right intervention, and recovering the money.
Example directions: Payment degradation → root cause → recovery action, Checkout drop-off recovery, Failed-subscription recovery, B2B receivables chaser, Mandate retry sequencer, Hinglish voice recovery, Promise-to-pay tracker.
The bar: Don’t just identify the problem. Show measured money recovered across a batch, with compliant escalation, stopping rules, and an audit trail.
Run the books and the cash position.
Build an agent that closes one finance-ops loop across a 50+ record batch of synthetic data, reporting its match rate and the exceptions it could not resolve.
Why now: The 2026 builder consensus: verification capacity, not generation speed, is the bottleneck. Reconciliation, settlement and forecasting are still done by hand.
Example directions: Multi-source reconciliation, Settlement Q&A agent, Forward cash forecaster, Tax-line matcher.
The bar: Throughput plus measured accuracy plus an honest exception list. One cherry-picked match proves nothing.
Build what you believe should exist.
Have an idea that doesn’t fit the tracks above? Build it. Pick a real problem, use AI meaningfully, and show us something that works. Any domain, workflow, or user is fair game.
Why now: The best ideas don’t always fit a predefined category. This track exists for builders who see an opportunity we didn’t.
Example directions: Surprise us, Solve a problem you deeply understand, Build something we haven’t thought of.
The bar: Open doesn’t mean easier. Show a real problem, a working product, meaningful use of AI, and evidence that it creates value. The same bar for execution, reliability, and depth applies here.
₹75,000 (monthly stipend) · 6 or 12 (months, your choice) · In-person (Bangalore, from September). Shortlisted builders go straight to a panel. No aptitude test. No group discussion.
Your code speaks louder than your resume.